A Methodology for Estimating Co-Scheduling Slowdowns due to Memory Bus Contention on Multicore Nodes

Abstract

When two or more programs are co-scheduled on the same multicore computer they might experience a slowdown due to the limited off-chip memory bandwidth. According to our measurements, this slowdown does not depend on the total bandwidth use in a simple way. One thing we observe is that a higher memory bandwidth usage will not always lead to a larger slowdown. This means that relying on bandwidth usage as input to a job scheduler might cause non-optimal scheduling of processes on multicore nodes in clusters, clouds, and grids. To guide scheduling decisions, we instead propose a slowdown based characterization approach. Real slowdowns are complex to measure due to the exponential number of experiments needed. Thus, we present a novel method for estimating the slowdown programs will experience when co-scheduled on the same computer. We evaluate the method by comparing the predictions made with real slowdown data and the often used memory bandwidth based method. This study show that a scheduler relying on slowdown based categorization makes fewer incorrect co-scheduling choices and the negative impact on program execution times is less than when using a bandwidth based categorization method.

Contention-aware application performance prediction for disaggregated memory systems

“…The rest of this section presents prior efforts which identify, quantify or model the applications' performance due to shared resource contention. Slowdown based methods -De Blanche et al [12,13] propose a slowdown based characterization method to estimate the applications' slowdown when sharing the memory bus. Their sensitivity curve is obtained using synthesized memory traffic and the contentiousness is found using performance counters.…”

“…Cache contention -Several works have been proposed to model how applications' performance is affected by changes in the amount of available shared resources, especially cache capacity. Our work does not consider using cache interference, as it is misleading when coscheduling applications using separate cache hierarchies [12].…”

“…Our work is differentiated from prior work [12,16,23], and provides improvements for a singular reason: Our approach targets performance prediction due to sharing of disaggregated memory, while the prior works use application working set size or local bandwidth as their measure of pressure to create the sensitivity curve. As noted in Section 3.4, cache contention characterization method is misleading for predicting the performance of applications using separated cache hierarchies.…”

“…In this section, we compare our methodology to the state-of-theart slowdown based methodology, Memgen [12]. Memgen uses a modified version of the Triad workload to generate a specific amount of traffic, then creating contention for the target application.…”

“…For this reason, coscheduling and resource allocation decisions for disaggregated memory require contention awareness in order to optimize application performance and overall system throughput. Slowdown based methods have been successful for single node coscheduling [12,16,23], but they have not been applied to disaggregated memory. This paper presents a slowdown based method for disaggregated memory.…”